On 20.11.23 08:58, Peter Eisentraut wrote:
> On 17.11.23 19:39, Paul Jungwirth wrote:
>> But I feel the overall approach is wrong: originally I used hardcoded
>> "=" and "&&" operators, and you asked me to look them up by strategy
>> number instead. But that leads to trouble with core gist types vs
>> btree_gist types. The core gist opclasses use RT*StrategyNumbers, but
>> btree_gist creates opclasses with BT*StrategyNumbers.
>
> Ouch.
>
> That also provides the answer to my question #2 here:
> https://www.postgresql.org/message-id/6f010a6e-8e20-658b-dc05-dc9033a694da%40eisentraut.org
>
> I don't have a good idea about this right now. Could we just change
> btree_gist perhaps? Do we need a new API for this somewhere?
After further thought, I think the right solution is to change
btree_gist (and probably also btree_gin) to use the common RT* strategy
numbers. The strategy numbers are the right interface to determine the
semantics of index AM operators. It's just that until now, nothing
external has needed this information from gist indexes (unlike btree,
hash), so it has been a free-for-all.
I don't see an ALTER OPERATOR CLASS command that could be used to
implement this. Maybe we could get away with a direct catalog UPDATE.
Or we need to make some DDL for this.
Alternatively, this could be the time to reconsider moving this into core.